Banks and regulation

Can a bank use an AVM under CRR3?

Yes. A bank may use a model to monitor the value of mortgage collateral and to identify which properties need a revaluation. CRR3 inserted Article 208(3a) into the Capital Requirements Regulation, which allows "advanced statistical or other mathematical methods" for both, under six conditions. The EBA guidelines also let a model update a value when no review trigger has fired (paragraph 226). A model may not value collateral alone when a loan is made (paragraphs 209 and 210), and it may not be the sole means of a revaluation after a possible material decline (paragraph 225).

How the permission got its conditions

From the time the CRR first applied in 2014, Article 208(3) ended with a single permission and no conditions: "Institutions may use statistical methods to monitor the value of the property and to identify property that needs revaluation."

The conditions came next, in guidelines. Section 7.4 of the EBA Guidelines on loan origination and monitoring (EBA/GL/2020/06), which have applied since 30 June 2021, sets out what a statistical model used for collateral must meet. Guidelines are not law. Supervisors and banks "must make every effort to comply", and each supervisor reports to EBA whether it does.

CRR3 then moved most of that wording into the regulation. From 1 January 2025 the old sentence is gone and a new paragraph 3a stands in its place. The permission is the same, and its conditions now sit in a regulation that applies directly in every Member State. The EBA guidelines still apply alongside it.

Where each condition comes from

Condition in CRR Article 208(3a)	Text	Same idea in EBA/GL/2020/06
Opening words	The methods are "developed independently from the credit decision process"	No equivalent in Section 7.4 (paragraph 231(d) asks the same of valuers)
(a) Policies	Criteria for using models, accounting for their "proven track record, property-specific variables considered, the use of minimum available and accurate information, and the models' uncertainty"	Paragraph 236, the same list
(b)(i) Granularity	"property- and location-specific at a sufficient level of granularity"	Paragraph 237(a), which gives "postcode" as the example
(b)(ii) Accuracy	"valid and accurate, and subject to robust and regular back-testing against the actual observed transaction prices"	Paragraph 237(b)
(b)(iii) Sample	"based on a sufficiently large and representative sample, based on observed transaction prices"	Paragraph 237(c)
(b)(iv) Data	"based on up-to-date data of high quality"	Paragraph 237(d)
(c) Responsibility	The institution is "ultimately responsible for the appropriateness and performance of the models"	Paragraph 238, which adds that the valuer stays responsible for a valuation made with a model
(d) Documentation	"the documentation of the models is up to date"	Paragraph 238
(e) Systems and data	"adequate IT processes, systems and capabilities" and "sufficient and accurate data"	Paragraph 239
(f) Validation	Model estimates are "independently validated", in a process "generally consistent with the principles set out in Article 185"	No equivalent in Section 7.4

Compared with the guidelines, the regulation adds two conditions, independence from the credit decision and independent validation. The rest moved up from a guideline into the regulation. The four model criteria of EBA paragraph 237 return almost word for word as points (b)(i) to (iv).

What a model may do at each stage of a loan

Stage	Can a model do it alone?	Where the rule is
Valuing collateral when the loan is made	No. A valuer with a full visit, inside and outside (paragraph 209). For residential property "in well-developed and mature property markets", a valuer's desktop valuation supported by a model with a confidence measure (paragraph 210)	EBA/GL/2020/06, paragraphs 209 and 210
Monitoring the value of the book	Yes, if the six conditions are met. The minimum is once a year for commercial and once every three years for residential property (Article 208(3)(a))	CRR Article 208(3a); EBA paragraphs 221 to 223
Picking the properties that need a revaluation	Yes, same conditions	CRR Article 208(3a)
Updating a value when no review trigger has fired	Yes: "either a revaluation carried out by a valuer or appropriate statistical models" that meet the Section 7.4 conditions	EBA paragraph 226
Reviewing a value after information points to a material decline	No. A qualified valuer independent of the credit decision; a model may support, but banks "should not use these models as the sole means of the revaluation"	CRR Article 208(3)(b); EBA paragraph 225
Loans above EUR 3 million or 5 % of own funds	A valuer reviews the valuation at least every three years	CRR Article 208(3)(b)
Setting the property value under Article 229(1)	Point (a) asks for an appraisal "by an independent valuer"; the value then reflects the monitoring results	CRR Article 229(1)

The last row leaves an open question. EBA paragraph 226 was written in 2020 against the old Article 229. Whether a model-only update now meets the independent valuer test in point (a) of the new Article 229(1) is not settled in the text, and banks should put the question to their supervisor. Our page on [CRR3 property value](/guides/crr3-property-value) sets out Article 229(1) in full.

The review trigger

Article 208(3)(b) sends a property to a valuer when information "indicates that the property value may have declined materially relative to general market prices". CRR3 added that "ESG-related considerations" count as such an indication.

EBA paragraph 224 asks banks to set out specific triggers in writing; its own example is "a change in the assumptions made in the appraisals". Three triggers a bank can write down for a model are:

1
The model or index shows a fall in value beyond a threshold the bank sets.
2
The model's confidence measure for that property is low.
3
Property-specific information casts doubt on the value: damage, a flood, a planning change, a poor energy rating.

The Hungarian central bank considers it good practice to set the threshold for the first trigger, a significant change in value, at no more than 20 percent (MNB Recommendation 15/2021, point 27). We cover the national rules in [national rules on AVMs in mortgage lending](/guides/avm-mortgage-rules-sk-cz-pl-hu).

What to ask an AVM vendor

Condition	Evidence to ask for	A warning sign
Independent development	Who builds and changes the model; whether anyone in credit can change an output	Values that can be edited per case without a log
Proven track record	Back-test reports with dates, sample sizes and the benchmark used	A single accuracy number with no date or sample
Property-specific variables	The list of inputs per valuation, and what happens when one is missing	Inputs the bank cannot supply at monitoring time, such as condition
Uncertainty	A confidence measure per value, and how it maps to an expected error spread	A confidence score that has never been tested against outcomes
Granularity	The smallest area the model distinguishes	Prices set per district or region only
Back-testing against transaction prices	The source of the transaction prices in your market, the period, and the result	Tests against asking prices presented as accuracy
Large, representative sample	Counts by region, property type, age and size	Strong results only where data is dense
Up-to-date, high-quality data	Data refresh frequency; how duplicates and errors are removed	No answer on duplicates
Documentation	Model description, version history, change log	Versions that change without notice
Validation	Who validates, how often, and the last report	Validation done by the people who built the model

The back-testing row is the hard one in Central Europe. Slovakia publishes no flat-by-flat sale prices, and the official statistics of Czechia and Hungary give averages only; Poland has published its register of transaction prices free of charge since 13 February 2026. Our page on [asking prices and sale prices in the four markets](/guides/asking-vs-transaction-prices-cee) lists what exists. A vendor in this region should say which benchmark it used, and a bank's own files of valuations and completed sales are often the best benchmark available.

The valuer's side

Paragraph 233 of the guidelines points the other way and asks banks to assess the accuracy of their valuers, "e.g. by backtesting on the value of the collateral through advanced statistical models". A model that checks a valuer's report against the market is a use the guidelines name directly.

Where Bytero Engine fits

Bytero Engine returns a value per property with a range; in the Bytero app it also shows the inputs and the comparables behind it. It can feed monitoring and the selection of properties for review, and it can serve as a second opinion on a valuer's report. It does not replace the valuer at origination or after a material decline, and the bank stays responsible for the model it uses. How a bank can test a vendor's figures is on [how AVM accuracy is measured](/guides/avm-accuracy-measurement).